Every AI time-savings number answers a question, and it is rarely the question you are asking. The vendor slide claiming four hours a week per person, the dashboard crediting your team with 1,400 assisted hours, the survey headline saying the average user saves 5.4%: all three can be honest, and none says how many hours came back to your company. Getting from a claim to a returned hour takes six questions, and most claims fail on the second.
Published figures differ by a factor of five or more, and that spread is not a disagreement about facts. It is a disagreement about what is counted, whose hours sit in the denominator, and whether the time spent fixing the output was subtracted.
What a time-savings number is counting
There are three ways anyone produces a time-savings figure, and they are not interchangeable.
Measured. Somebody timed a task with the tool and without it. The only method that yields a difference rather than an estimate, and rare outside research settings because it needs a control group or a recorded before-state.
Reported. Somebody asked people how much time they saved. The most careful version is the question the Federal Reserve Bank of St. Louis uses in its Real-Time Population Survey: how many additional hours would you have needed last week to do the same work without the tool. Still a counterfactual nobody can observe.
Imputed. Somebody counted events and assigned each one a time value. This is the method behind most enterprise dashboards and the one buyers least often recognize, because the output arrives in hours and looks measured.
Microsoft is unusually open about doing this. Its published method for Copilot Assisted Hours credits six minutes per information search, six minutes per piece of content created from scratch, and a full hour for summarizing an hourlong meeting. Read the schedule and the number becomes arguable: if the summary gets skimmed and the meeting still runs its full hour, that hour was credited, not saved.
Imputation is not dishonest, but a credit schedule can be recomputed, which is the first move to make with any dashboard total.
Six questions that decode any claim
Put these to any figure, in order. The first one it fails is where the conversation goes.
| # | Question | What a weak answer looks like |
|---|---|---|
| 1 | What task? Which single unit of work, start to finish? | “Knowledge work,” “admin,” “general productivity” |
| 2 | Whose hours? Users only, or everyone holding a license? | A per-person figure with no stated population |
| 3 | Measured, reported, or imputed? | Hours presented with no method named |
| 4 | Gross or net? Are prompting, review and rework inside the boundary? | Savings on the first draft, silence on the second |
| 5 | Against what? A baseline, a comparison group, or a memory? | “Before AI we used to spend about…” |
| 6 | Where did the hour go? Which deliverable, backlog or overtime line moved? | “Capacity was freed up” |
Question six is the one almost nothing in the market answers and the only one a CFO cares about. An hour saved that nobody reassigned is not a cost reduction. It is a slightly easier Tuesday. Quality sits beside all six: a speed figure with no quality score beside it is the easiest number here to produce and to dismiss.
An honest claim names its own answers: on the weekly client update, for the 31 people who write one, timed against a baseline from the week before training, net of review, median time fell from 95 to 74 minutes. Almost nothing published reads like that.
The published claims, decoded
The same decode applied to the figures you are most likely to be handed. They are not competing answers to one question.
| Claim | Method | Population | What it supports |
|---|---|---|---|
| 56% faster on a coding task (Peng et al.) | Measured, randomized | Recruited developers, one standardized task | A ceiling on clean, self-contained work |
| 19% slower (METR) | Measured, randomized | 16 developers, 246 issues, familiar repositories | Complex familiar work can slow down |
| 40% faster writing, quality up 18% (Noy and Zhang) | Measured, randomized | 453 professionals | Short writing tasks compress a lot |
| 14% more resolutions, 34% for the newest staff (Brynjolfsson, Li and Raymond) | Measured in the field | Support agents at one firm | Gains concentrate in newer staff |
| 5.4% of work hours saved (St. Louis Fed) | Reported counterfactual | Weekly users only | What users believe about their own week |
| About 1.4% of total work hours, rising toward 2.2% | Reported counterfactual | All employed adults | The economy-wide figure, non-users included |
| 11 minutes a day per user | Imputed from a credit schedule | Active users of one product | A usage threshold, not a measured hour |
| 37% of saved time lost to rework (Workday) | Reported, 3,200 respondents | Mixed | The discount to apply |
Lab results on clean tasks are a ceiling, not a forecast. Field results on complex work can be negative. And the two most-quoted enterprise figures, per-user minutes and the economy-wide percentage, measure such different things that neither checks the other.
Why the numbers disagree by five times
A vendor claiming four hours a week is claiming about 10% of a 40-hour week. A national survey reporting roughly 2% of total work hours is five times below that. Both can be right, for three reasons that compound.
The denominator. Vendor figures are per active user. The Fed’s low number spans all employed adults, most of whom do not use these tools weekly. Among weekly users it rises to 5.4% of hours, which closes over half the gap by itself.
The boundary. A credit schedule stops when the output appears. The Workday survey found 37% of saved time consumed by fixing low-quality output, only 14% of employees reporting consistently net-positive outcomes, and the heaviest users losing around a week and a half a year to correction.
Selection. People who stay in a study are not a random sample. METR hit this directly: repeating its 2025 trial a year later, developers would not give up their tools even temporarily, the control arm collapsed, and the follow-up produced intervals its own authors called weak evidence.
External numbers are priors, not substitutes. Use the published range to judge whether your own result is plausible, then produce one, by the method set out in our guide to proving capability moved.
Gross hours, net hours, returned hours
Three quantities, routinely reported as one. Written as separate lines, the argument becomes visible.
| Line | Definition | Usual treatment |
|---|---|---|
| Gross hours | Task time before minus task time after | The headline |
| Minus prompting and iteration | Getting to a usable output | Inside your number if you timed the whole task |
| Minus review and rework | Checking and fixing | Usually omitted; apply a measured discount |
| Minus tool overhead | Onboarding, administration, troubleshooting | Almost never counted |
| Net hours | What the person genuinely has spare | Rarely published |
| Minus unreallocated time | Net hours that went nowhere named | Never counted |
| Returned hours | Hours that landed on a named output | The only line a CFO can bank |
The step from net to returned decides whether a program shows up in the business. A team that shaves 90 minutes off report preparation and spends it on more thorough report preparation has a better report, not a returned hour. The same team taking on a second client with no new headcount has returned 90 minutes.
So the landing place is chosen in advance: a deliverable produced more often, a backlog down by a stated amount, an overtime line that falls, a requisition not filed. Name it before the training, alongside the before-figures in our look at what to measure the week before training starts. Chosen afterwards, it becomes whichever number happened to move.
The perception gap runs both ways
The most-quoted finding in this field is METR’s 2025 trial, and it is usually quoted wrong. Sixteen experienced developers worked 246 issues in large repositories they already knew. With AI tools allowed they took 19% longer. They had forecast a 24% speedup, and after living through the slowdown still believed the tools had made them about 20% faster.
That gets cited as proof people overstate gains. The opposite error is equally documented: in the randomized Copilot study, participants finished 56% faster and put their own gain at about 35%, understating it by 20 points.
Self-report is uncalibrated, then, rather than inflated, which has a more useful implication. Keep asking, but collect the estimate separately from the objective timing and treat the difference as its own finding, at a cost of one survey question beside one timed deliverable. In METR’s 2026 survey of 349 technical workers, the subgroup reporting the lowest gains was its own research staff, the people most aware of the bias. Calibration is learnable, and it moves estimates down.
A protocol for measuring hours returned
What a company of 50 to 500 people can run without an analytics function, in five steps.
- Pick one unit of work. Produced at least weekly, by a group you can name, with a visible start and end. A client update, a vendor invoice, a candidate screen, a close pack. Not “email.”
- Time it before, twice. Two separate weeks, same group, recorded at the time rather than recalled. Timestamps from systems you already run beat a stopwatch: ticket open and close, document created and shared, first draft and final.
- Define the boundary out loud. Start when the work is picked up, stop when whoever receives it accepts it. Prompting, review and rework then sit inside the window by construction, so the discount never has to be argued.
- Hold one group back for six weeks. A comparison group is the cheapest credibility available and need not be a true control. Train one team in March and a comparable team in May, then compare the same deliverable on the same dates. It also separates your training from a tool release nobody told you about, a confound our note on the 30/60/90 checkpoints treats as a gate.
- Report the hour and its landing place together. One line for net hours per person per week, one for what absorbed them, one for a quality score on the same deliverable. No composite.
Price the hours with a published figure rather than salary alone. The Bureau of Labor Statistics put employer costs for civilian workers at $49.46 per hour worked in its June 2026 reference period, $33.85 of it wages and salaries and $15.61 benefits. Using total compensation, and saying so, removes one of the two objections finance will raise. The other is reallocation, which is what line three is for.
What the measurement costs in hours
Method guides never state this, which is why they get adopted in January and dropped in March. Per group of 50, per quarter:
- Choosing the unit of work and the landing place: 3 hours, mostly argument
- Two before-weeks of timing, pulled from existing systems: 4 hours
- One after-week on the same basis: 2 hours
- Quality scoring, ten minutes per person on one real sample: 8 hours of manager time
- The perception question and the write-up: 3 hours
About 20 hours a quarter, two-thirds of it manager rather than analyst time, which is the reason to measure one deliverable properly instead of six badly. Scoring the same group’s capability is a separate instrument, set out in our framework for scoring a team before and after.
A worked decode
A 240-person services company is handed a vendor claim of 4.2 hours saved per person per week. Figures are illustrative; the arithmetic is the point.
| Step | Working | Running figure |
|---|---|---|
| Vendor claim, per active user | As presented | 4.2 hrs/week |
| Population: 58% of licensed staff are weekly users | 4.2 × 0.58 | 2.4 hrs/week |
| Method: imputed from a credit schedule, not timed | Halve pending local timing | 1.2 hrs/week |
| Rework, at the surveyed 37% | 1.2 × 0.63 | 0.76 hrs/week |
| Local timing on one deliverable, net | Measured | 0.9 hrs/week |
| Reallocation: 60% landed on a named second deliverable | 0.9 × 0.6 | 0.54 hrs/week returned |
| Annualized across 240 staff at $49.46 | 0.54 × 52 × 240 × 49.46 | about $333,000 |
The claim shrank eightfold and became defensible. Local timing landed slightly above the discounted vendor figure, the normal result when the imputation was conservative and the training was real. Reallocation is the only step a reader can check without your spreadsheet, which is why it belongs in the report.
Licenses arrive on a purchase order and capability does not, the argument in our case that buying licenses is not the same as building capability. An hours figure that survives the six questions tells you which one you bought.
Frequently asked questions
How do you measure time saved by AI?
Time one complete unit of work before and after, with prompting, review and rework inside the boundary, against a group not yet trained. Then state where the saved hours landed. Skip either and you have a gross figure finance will not accept.
Can we just ask employees how many hours they save?
Ask, but do not stop there. Self-report is uncalibrated rather than inflated: developers in one randomized trial believed they were 20% faster while measuring 19% slower, and participants in another understated a real 56% gain as 35%. Collect the estimate beside an objective timing and treat the gap as a finding.
Why do AI time-savings statistics vary so much?
They count different things. Lab studies time one clean task, surveys ask about a remembered week, dashboards impute hours from event credits, and economy-wide figures divide by a workforce that mostly does not use the tools. Check population and method before comparing two.
What is a realistic time saving per employee?
For a trained group using these tools weekly, the reported range centers near 5% of work hours, about two hours a week, and less net of rework. Treat anything above four hours a week as a claim needing a stated method, not a benchmark.
Does dashboard reporting count as measurement?
Not on its own. Most dashboard totals are imputed from a credit schedule, six minutes per information search for instance, and the schedule is usually published. Use the dashboard for reach and habit; use timing for hours.
What should we subtract from a gross hours figure?
Prompting and iteration, review and rework, tool overhead. Time the whole workflow end to end and the first two are already inside your number. For rework, the surveyed 37% is a fair placeholder until you have your own.
How do we prove the saved hours were used?
Choose the landing place before the training and measure it directly: more units of the same deliverable, a backlog down by a stated amount, a fall in overtime, a role left unfilled. If none moved, report net hours and say reallocation has not happened.
Our measured savings are smaller than the vendor claim. Did the training fail?
Usually not. A vendor figure is per active user, often imputed, almost always gross. After adjusting for weekly-user share and rework, a credible local result often lands at an eighth to a quarter of the headline. Judge against your own baseline, not the slide.
Key takeaways
- Every time-savings number is measured, reported or imputed. Establish which before repeating it.
- Six questions decode any claim: what task, whose hours, which method, gross or net, against what, and where the hour went. Quality is scored alongside all six.
- Vendor per-user minutes and economy-wide percentages differ fivefold or more because the denominators differ. Neither checks the other.
- Only returned hours reach the business, and only if the landing place was named in advance.
- Self-report is uncalibrated in both directions. Collect it beside objective timing and report the gap.
- Time the whole workflow so review and rework sit inside the boundary, price hours at $49.46 of total employer compensation, and budget about 20 hours a quarter per group of 50.
If you want this run against one of your own deliverables, with before-figures recorded and a landing place named before anyone sits down, book a free AI readiness call. We will tell you which unit of work to time, what to subtract, and what the number will not prove.
Arthur Wandzel